Banca de DEFESA: EDUARDO JACOMO SERAPHIM NOGUEIRA

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : EDUARDO JACOMO SERAPHIM NOGUEIRA
DATE: 29/08/2025
TIME: 19:00
LOCAL: a definir
TITLE:

Forecasting Central Government Primary Expenditure: A Comparative Analysis of Econometric Techniques, Machine Learning, Deep Learning, and Forecast Combination..

KEY WORDS:

forecasting. time series. fiscal policy.

PAGES: 80
BIG AREA: Ciências Sociais Aplicadas
AREA: Economia
SUMMARY:

The forecasting of public expenditures is essential for fiscal planning and the sustainability

of government accounts. However, many countries, especially those with low and

middle incomes, still rely on subjective methods and simple spreadsheet extrapolations
for fiscal forecasts. Despite the growing use of traditional econometric methods, the
application of machine learning and deep learning techniques remains limited and largely
concentrated on forecasting public revenues. International studies show gains from using
machine learning and deep learning algorithms, particularly due to their ability to handle
nonlinearities and complex patterns. However, robust evidence is still lacking for short
and noisy monthly fiscal series, which are common in emerging countries. In the Brazilian

context, there is a scarcity of research focused on disaggregated forecasting of federal
primary expenditures and systematic comparisons across different classes of models. This
work empirically investigates the predictive performance of traditional models, machine
learning models, deep learning models, and model combinations in forecasting monthly
series of Brazilian federal expenditures. To this end, we combined statistical benchmarks
and supervised algorithms, performing automatic hyperparameter optimization to obtain
robust point forecasts. We used temporal cross-validation as a model selection procedure
and conformal prediction to generate calibrated confidence intervals. Various error metrics

were employed to compare the performance of the approaches. Overall, statistical
models proved highly competitive, outperforming machine learning and deep learning

algorithms. Although inter-class ensembles did not minimize average errors, they increased
the robustness of the forecasts. Based on official data from the Federal Government, this
study demonstrated that time series forecasting techniques are an important tool to

support the work of fiscal policy makers. Future research may explore exogenous variables
and structural or semi-structural models to enhance the accuracy, robustness, and

interpretability of predictive models.

 

COMMITTEE MEMBERS:
Interno - 1642911 - DANIEL OLIVEIRA CAJUEIRO
Externo ao Programa - 1165891 - JOAO GABRIEL DE MORAES SOUZA - nullInterno - 1550794 - JOSE GUILHERME DE LARA RESENDE
Externo à Instituição - JOÃO FROIS CALDEIRA - UFSC